Upload
lamkiet
View
239
Download
0
Embed Size (px)
Citation preview
Emilio Ancillotti, Raffaele Bruno, Marco Conti
IIT – CNR, Pisa, Italy
Speaker: Chiara Boldrini
Questions to: [email protected]
PE-WASUN 2008
Outline o Why rate adaptation?
o General approaches for rate adaptation in 802.11 networks
o Rate adaptation algorithms in MadWifi driver • AMRR
• ONOE
• SampleRate
o Indoor experiments
o Outdoor experiments
o Conclusions and ongoing work
2
IEEE 802.11 Multi-rate capabilities o The 802.11 a/b/g standards allow the use of
multiple transmission rates • 802.11b, 4 rate options (1,2,5.5,11Mbps) • 802.11a, 8 rate options (6,9,12,18,24,36,48,54 Mbps) • 802.11g, 12 rate options (11a set + 11b set)
o Different bit rates are provided by employing different modulation schemes and coding rates
o Rate Adaptation refers to the algorithms used to select the transmission rate that provides the best “link performance”
o Rate adaptation plays a critical role to the throughput performance, but it is yet unspecified by the 802.11 standards
3
Motivations for rate adaptation o The link-layer capacity at each data rate depends on channel
quality, as well as various environmental dynamics, such as: • Channel fluctuations • Node mobility • Medium contention
o In practical settings, wireless channels can be extremely dynamic (e.g., due to multipath fading)
Theoretical MAC layer throughput for an AWGN
channel versus SNR (payload:2000Byte,
802.11a modes)
Rate adaptation: general approaches
o Signal strength-based algorithms: • Rate adaptation relies on physical layer measurements (SNR,
RSSI) • They require an accurate channel model • Generally not compliant with the 80211 standards (e.g.,
RBAR and OAR)
o Statistics-based algorithms: • Rate adaptation relies on frame transmission statistics (e.g.,
number of retries, number of consecutive frame successes or failures)
• Rate is decreased upon severe loss • Generally probe packets are used to create long-term
statistics • Need to differentiate between collisions and channel errors • Several examples: ARF, AMRR, ONOE, SampleRate, CARA
Experimental evaluation in real scenarios
o Schemes compliant with 802.11 standards have been implemented in commodity hardware and open software drivers
o Practical investigations of rate adaptation performance o Available studies have mainly focused on indoor wireless
networks, considering: • the impact of channel dynamics due to rapid fluctuations of the receive
signal strength • random channel errors, mobility-induced channel variations, and
contention from hidden stations
o Experimental studies have been conducted mostly in small wireless networks consisting of an AP and a few clients
o How these autorate algorithms cope with moderate to high medium contention levels?
o How these autorate algorithms perform on medium-distance 802.11 links?
7
MadWifi: Multiple Rate Retry (MRR)
o MadWifi driver enables the network interface to transmit at different data rates the retransmissions of a given frame
o Four rates (r0,r1,r2,r3) and transmission counts (c0,c1,c2,c3) are associated to each frame
o Each rate ri is tried ci-times before using next rate
o c0+c1+c2+c3 is the maximum number of allowed retransmissions (<=ATH_TXMAXTRY)
o Several rate adaptation algorithms employ the multiple rate retry capabilities of the MadWifi driver
8
Adaptive Multiple Rate Retry (AMRR)*
o AMRR sets c0=c1=c2=c3=1, i.e., each rate is tried just once
o Rate r3 is set to the lowest bit rate (1Mbps in 11b/g and 6Mbps in 11a)
o An heuristic is used to select r0:
• If less than 10% of the packet transmissions failed during the last observation period, then increase the data rate
• If more than 33% of the packet transmissions failed during the last observation period, then decrease the data rate
o Rate r1, is the rate immediately lower than r0, and rate r2 is the rate immediately lower than r1
*M. Lacage, M. H. Manshaei, and T. Turletti. IEEE 802.11 Rate Adaptation: A Practical Approach. In Proc. MSWiM ’04, pages 126–134, Venice, Italy, 2004.
9
ONOE* o ONOE algorithm is a variant of the AMRR scheme
o ONOE uses larger retransmission counts than AMRR (c0=4,c1=c2=c3=2)
o ONOE sets r1, r2, r3 bit rates as AMRR
o An credit-based heuristic is used to select r0:
• If less than 10% of the packet transmissions failed during the last observation period, then the credits of r0 are increased by one; otherwise the credits of r0 are reduced by one
• If more than 10% of the packet transmissions failed during the last observation period, then the credits of r0 are reduced by one
• If r0 has more than 10 credits, then increase the data rate
• If more than 50% of the packet transmissions failed during the last observation period, then decrease r0
*MadWifi driver documentation. Onoe Rate Control. http://madwifi.org/wiki/UserDocs/RateControl.
10
SampleRate*
o SampleRate estimates the medium contention level by evaluating the expected transmission time for a frame at different data rates
o SampleRate transmits each frame at the rate r that has the shortest expected transmission time
o A probe packet at a different rate is sent every ten frames
o SampleRate probes only rates with a minimum packet transmission time (i.e., with n=0) lower than the average transmission time of the current bit rate
*J. Bicket. Bit-rate Selection in Wireless Networks, February, 2005. Massachusetts Institute of Technology, M.S. Thesis.
€
tx _ time(r,n,L) = backoff (n +1) + (n +1) ⋅ (Δ + L /r)
MadWiFi: frame transmission Associate a new frame with a transmission rate, which is estimated after the completion of the previous ath_rate_tx_complete() function call
11
Set the properties of the transmit process descriptor
The frame is enqueued into the ath_txq, which is hardware queue in the MAC controller
Process the completed transmit descriptors
ath_hal_setuptxdesc(ah, ds , pktlen/* packet length */ , hdrlen/* header length */ , atype/* Atheros packet type */ , MIN(ni->ni_txpower, 60)/* txpower */ , txrate, try0/* series 0 rate/tries */, antenna/* antenna mode */ , ctsrate/* rts/cts rate */ , ctsduration/* rts/cts duration */…);
12
Indoor experiments
o Indoor experiments aim at evaluating the impact of contention level on rate adaptation performance
o Hardaware/software setup: • 12-node network composed of IBM Thinkpad model R50E
laptops • Each node has one NetGear WPN511 card operating on channel
11 in 802.11g mode • MadWiFi driver version 0.9.4
o Traffic configurations • UDP traffic generated with iperf (packet size 1500Bytes)
• Single-hop flows
• Each test lasts 2 minutes and is repeated five times
Indoor Experiments: Throughput
13
o Maximum throughput achieved with fixed rate at 54Mbps
o Considerable throughput degradation when autorate is used. With 11 saturated stations the throughput achieved with loss ratio threshold-based schemes (i.e., AMRR and ONOE) can be up to ten times lower than the best throughput obtained with a fixed transmission rate
Indoor Experiments: retry and rate distributions
14
o AMRR adopts small retry limits, which induces high rate variability
o Both AMRR and ONOE operate to keep the frame loss rate below pre-determined and fixed thresholds (33% for AMRR and 50% for ONOE)
o SampleRate overestimates the maximum throughput achievable at the different transmission rates, leading to a conservative rate selection
o Channel quality is good
o Contention induces collisions
o More retries with higher data rates and several stations: synchronization problems?
15
Outdoor Eperiments
280m 170m
190m
280m
160m
Directional links Omnidirectional links
A
B
C
D
E
Mesh router o Soekris net4801 router
• 266 MHz NSC SC1100 single chip processor
• 256 Mbyte SDRAM, soldered on board
• 100G 2.5" Hard Drive • 1-3 10/100 Mbit Ethernet ports,
RJ-45 • USB 1.1 interface • Mini-PCI type III socket • PCI Slot • Two Atheros AR5414 miniPCI
modules, 20dBm/100mW, Wireless Super AG, 802.11a/b/g/108Mbps 5/2.4GHz
• Omni-directional and directional antennas usable in the 2.4GHz band
Soekris Board
16
Outdoor experiments
17
o Frame loss rates are not uniform over all transmission rates*: • Transmission rates with negligible
loss rates • Transmission rates not working at all • Transmission rates with intermediate
loss rates
o Allan deviation used to measure the correlation of channel errors.
o Xi is the average loss rate over a time interval T
€
dev =
xi − xi−1( )2i=1
n
∑2n
*G.. Bianchi, F. Formisano, D. Giustiniano. 802.11b/g Link Level Measurements for an Outdoor Wireless Campus Network. In Proc. WoWMoM’06, pages 525–530, Niagara-Falls, NY, 2006.
Outdoor experiments: throughput
18
o In our mesh network, the are transmission rates with negligible frame loss ratios.
o These links are also quite stable, and observation periods of one second give reliable estimates of the long-term average frame loss rate
o Rate adaptation schemes based on loss ratio thresholds will perform reasonable well on these links
Outdoor experiments: rate distribution
19
o AMRR works reasonably well in both gradual and steep links, but it is worse than SampleRate:
• AMRR permits only four consecutive retries for each frame • AMRR occasionally tries very low transmission rate
o ONOE works reasonably well for steep links (even better than AMRR), but is the worst for gradual links:
• ONOE is very slow in increasing the rate
o SampleRate seems the best for static configurations with non-bursty links transmission rate
20
Design guidelines for congestion-aware rate adaptation
o We need more accurate techniques to correctly estimate the medium contention level • How to compute the correct time interval between successful
transmissions?
o We should make the sampling period used to estimate the long-term frame loss rates adaptive to channel temporal correlations • How to measure the channel temporal correlation?
o Minimize the use of probe packets for low transmission rates
o Ongoing activity • Design of accurate 802.11 compliant strategies to differentiate
collisions from channel errors
• Extensions of SampleRate to correctly operate in collision dominated environments.